SLIDE 1
Hate Speech Detection is Not as Easy as You May Think: A Closer Look at Model Validation
Aymé Arango, Jorge Pérez and Bárbara Poblete
UNDETECTED HATE SPEECH IN SOCIAL MEDIA
VS
ALMOST PERFECT STATE-OF-THE-ART RESULTS
Hate Speech Detection is Not as Easy as You May Think: A Closer Look - - PowerPoint PPT Presentation
Hate Speech Detection is Not as Easy as You May Think: A Closer Look - - PowerPoint PPT Presentation
Hate Speech Detection is Not as Easy as You May Think: A Closer Look at Model Validation Aym Arango, Jorge Prez and Brbara Poblete UNDETECTED ALMOST PERFECT HATE SPEECH VS STATE-OF-THE-ART IN RESULTS SOCIAL MEDIA UNDETECTED HATE
SLIDE 2
SLIDE 3
UNDETECTED HATE SPEECH IN SOCIAL MEDIA
SLIDE 4
ALMOST PERFECT STATE-OF-THE-ART RESULTS
94% F1
[Agrawal and Awekar] ECIR
2018
93% F1
[Badjatiya et al.] WWW
2017
92% F1
[Zeerak Waseem] NAACL
2016
SLIDE 5
Hate Speech Detection is Not as Easy as You May Think
We show that state of the art results are highly overestimated due to experimental issues in the models:
Including the testing set during training phase Oversampling the data before splitting User-biased datasets
SLIDE 6
State-of-the-art replication User distribution Generalization
SLIDE 7
State-of-the-art replication User distribution Generalization
SLIDE 8
ALMOST PERFECT STATE-OF-THE-ART RESULTS
94% F1
[Agrawal and Awekar] ECIR
2018
93% F1
[Badjatiya et al.] WWW
2017
92% F1
[Zeerak Waseem] NAACL
2016
SLIDE 9
DATASET 1
[Waseem and Hovy] NAACL 2016
Tweet Label Hate Non-Hate
SLIDE 10
Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction PHASE 2 Classification Method DATASET 1
[Waseem and Hovy] NAACL 2016
93% F1
SLIDE 11
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 12
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 13
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
TEST TRAIN Splitting
Embeddings Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 14
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
TEST TRAIN Splitting
AVG(Embeddings) GBDT Prediction
93% F1
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 15
This looks great! But there is a problem.
SLIDE 16
AVG(Embeddings) GBDT Prediction
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
TEST TRAIN TEST
Embeddings LSTM Softmax Prediction Fully Connected
Splitting
SLIDE 17
Let’s create the model only with the training set.
SLIDE 18
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
PHASE 1 Feature Extraction DATASET 1
[Waseem and Hovy] NAACL 2016
SLIDE 19
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
TEST TRAIN TEST TRAIN Same Splitting New PHASE 1 Feature Extraction
SLIDE 20
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
New PHASE 1 Feature Extraction TRAIN TEST TRAIN Same Splitting
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 21
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
New PHASE 1 Feature Extraction TEST TRAIN
Embeddings
TRAIN Same Splitting
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 22
PHASE 2 Classification Method Model 1
[Badjatiya et al.]
2017
New PHASE 1 Feature Extraction TEST TRAIN
73% F1
TRAIN Same Splitting
93% F1
Embeddings LSTM Softmax Prediction Fully Connected AVG(Embeddings) GBDT Prediction
SLIDE 23
The result is overestimated due to the inclusion of the testing set during the training phase.
SLIDE 24
Model 2
[Agrawal and Awekar]
2018
Oversampling Data Feature Extraction + Classification Method DATASET 1
[Waseem and Hovy] NAACL 2016
94% F1
SLIDE 25
DATASET 1
[Waseem and Hovy] NAACL 2016
Model 2
[Agrawal and Awekar]
2018
SLIDE 26
Oversampling Model 2
[Agrawal and Awekar]
2018
94% F1
Splitting TRAIN TEST
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 27
This also looks great! But there is another problem.
SLIDE 28
DATASET 1
[Waseem and Hovy] NAACL 2016
Model 2
[Agrawal and Awekar]
2018
SLIDE 29
Model 2
[Agrawal and Awekar]
2018
Oversampling Splitting TRAIN TEST
SLIDE 30
Splitting
79% F1
Model 2
[Agrawal and Awekar]
2018
94% F1
Oversampling TEST
Embeddings LSTM Softmax Prediction Fully Connected
SLIDE 31
The result is overestimated due to the fact that the
- versampling phase occurs before splitting the data.
SLIDE 32
However, there is another issue to take into account.
SLIDE 33
State-of-the-art replication User distribution Generalization
SLIDE 34
% Tweets from the most prolific user per class
96% 44% 38% 25%
96 % 44 % 25 %
Hate Non-Hate
Sexism Racism
SLIDE 35
TEST TRAIN Splitting without
- verlapped users
DATASET 1
[Waseem and Hovy] NAACL 2016
Model 1
[Badjatiya et al.]
2017
73% F1 93% F1 44% F1
Model 2
[Agrawal and Awekar]
2018
79% F1 94% F1 35% F1
SLIDE 36
What happens if we have a dataset with a better user distribution?
SLIDE 37
DATASET 1
250 tweets per user per class
NEW DATASET DATASET 2
[Davidson et al.] ICWSM 2017
DATASET 2
Hateful tweets
SLIDE 38
NEW DATASET TEST TRAIN Splitting without
- verlapped users
Model 1
[Badjatiya et al.]
2017
73% F1 93% F1 44% F1 78% F1
Model 2
[Agrawal and Awekar]
2018
79% F1 94% F1 35% F1 76% F1
SLIDE 39
User distribution on datasets has an impact
- n the classification results.
SLIDE 40
State-of-the-art replication User distribution Generalization
SLIDE 41
TRAINING SET TESTING SET
SLIDE 42
TRAINING SET DATASET 3
[Basile et al.] SemEval 2019
SLIDE 43
DATASET 1
[Waseem and Hovy] NAACL 2016
DATASET 3
[Basile et al.] SemEval 2019
Model 1
[Badjatiya et al.]
2017
47% F1
NEW DATASET DATASET 3
[Basile et al.] SemEval 2019
51% F1
DATASET 1
[Waseem and Hovy] NAACL 2016
DATASET 3
[Basile et al.] SemEval 2019
Model 2
[Agrawal and Awekar]
2018
51% F1
NEW DATASET DATASET 3
[Basile et al.] SemEval 2019
54% F1
SLIDE 44
Better user-distributed datasets lead to better generalization.
SLIDE 45
Conclusions
SLIDE 46
Hate Speech Detection is Not as Easy as You May Think
We show that state of the art results are highly overestimated due to experimental issues in the models:
Including the testing set during training phase Oversampling the data before splitting User-biased datasets
SLIDE 47
Hate Speech Detection is Not as Easy as You May Think: A Closer Look at Model Validation
Aymé Arango, Jorge Pérez and Bárbara Poblete